Learning by Doing and the Demand for Advanced
Products
Yufeng Huang∗
Tilburg University
October 7, 2014
Abstract
In markets where product usage requires consumer skills, learning by doing – the evolution of
these skills through using the product – plays an important role in shaping demand evolution
in the long run. In this paper, I measure the returns to experience on a sample of users of
digital cameras, via a measure of their picture quality. I then relate this measure to data on
camera usage and switching, and estimate a dynamic demand model with learning by doing. I
find that a 1-year increase in experience enhances a starting consumer’s camera-usage skills,
which raises both her welfare by an amount equal to her lifetime spending on cameras, and
her demand for advanced products in the same brand by 26%. Thus, not only is the provision
of knowledge valued by the consumer, but it also increases her product demand. On the other
hand, for some consumers, up to 26% of her experience is not transferable to products with
different designs – in particular across brands. Hence, learning from using products in this
category makes the consumer increasingly brand loyal.
Keywords: learning by doing, switching cost, dynamic programming, forward-looking behavior, digital
camera market
∗
I am grateful to my advisors, Bart Bronnenberg and Tobias Klein, for their constant guidance and support. I thank
MSI and two anonymous referees for their acknowledgment and detailed comments, in the 2013 Alden G. Clayton
Dissertation Competition. This paper also benefits from comments and suggestions from Jaap Abbring, Pradeep
Chintagunta, Marnik Dekimpe, Gautam Gowrisankaran, Jonne Guyt, Carlos Santos, Emanuele Tarantino and Yifan
Yu, and seminar/conference participants in Tilburg, Mannheim and the 2013 CEPR/JIE School on Applied IO. I thank
The Netherlands Organization for Scientific Research (NWO Vici Grant 453-09-004) for financial support. All errors
in this paper are mine.
1
1
Introduction
“I’d get a DSLR based upon my experience level. [...] If your situation is different to
mine however. [...] you’ll probably be quite happy with a cheaper point and shoot.”
– Darren Rowse,1 Should you buy a DSLR or Point and Shoot Digital Camera?
“Nikon and Canon are as good as each other overall. [...] The differences lie in
ergonomics and how well each camera handles, which is what allows you to get your
photo – or miss it forever. [...] and I can’t for the life of me figure out the menus of the
Nikon Coolpix cameras.”
– Ken Rockwell,2 Nikon vs. Canon
The two quotes above demonstrate a widely-held belief among practitioners – that novices and
expert consumers demand products of different quality; and among the experts, their demand is
specialized, and thus dependent on their previous experience. The role of consumer product experience is not unique to the digital camera industry. For products such as home electronics,
sports equipment, entertainment, food and beverages,3 marketing practitioners have long realized
the wide differences in the demand between novices and experts, and have targeted their different
needs by developing portfolios of differentiated products.
Despite the importance of experience accumulation in consumer demand, the quantitative understanding of this is limited. This is because standard choice data alone confound the returns to
experience with alternative explanations, such as changes in tastes or increases in awareness. In the
context of choices of digital cameras, this paper utilizes a unique data-set that provides a measure
1 Extracted from the following URL by March 2014. http://digital-photography-school.com/should-you-buy-adslr-or-point-and-shoot-digital-camera
2 Extracted from the following URL by March 2014. http://www.kenrockwell.com/tech/nikon-vs-canon.htm
3 Alba and Hutchinson (1987) are among the first to conceptualize the role of consumer product experience – “expertise”. In two experimental studies, Nam et al. (2012) and Clarkson et al. (2013) document the differences between
expert and novice consumers, in their choices of, respectively, digital cameras and food/beverages. Albuquerque and
Nevskaya (2012) model a consumer’s progressively higher tendency to play video games. Youn et al. (2008) document
that beginner climbers tend to choose entry level climbing gears, and will later progress into advanced but specialized
products.
2
of the returns to consumer experience, and quantifies its role in the demand for entry-level and advanced digital cameras. This allows for a better understanding of the long-run evolution of demand
from entry-level to advanced products, and potentially, better quantitative marketing decisions.
In this paper, a consumer of digital cameras cares about her picture quality, which she learns
to produce through the accumulation of experience. Hence, being able to measure the effect of
experience on picture quality, and the effect of increasing picture quality on her demand for advanced products, is key to understanding the effect of experience on demand in this case. For
this purpose, I collect individual panel data from pictures displayed on a photo-sharing website,
Flickr.com. And I exploit the fact that pictures are sorted by the date of upload, and I compare the
number of views among pictures that a consumer uploads at the same time. Since these pictures
are displayed together and are likely to be viewed together, the differences in views are more likely
to reflect picture quality differences. In addition, I observe variations in when and by which camera
each picture was taken, within the same batch of upload, and hence can infer the causal effect of
experience and equipment on the picture quality.
With up to 10 years of measurement of picture quality per individual, jointly with observations
of camera usage and switching,4 the role of experience accumulation is evident even without a
(structural) model. On the one hand, with their experience accumulating, consumers are capable
of producing higher quality pictures. On the other hand, after a consumer switches cameras, she
cannot immediately produce pictures of as high quality as she did with the previous one; and this
gap is larger for consumers with more experience. This indicates that not only is the consumer
obtaining general experience in photography, she is also accumulating specific knowledge about
using the given product.
To quantify the role of experience on demand in this context, I then construct a structural model
of a consumer’s demand for cameras and choices of product usage. In the model, the quality of
the camera that the consumer owns is complemented by her ability to use it – her “human capital”,
which improves with previous experience through learning by doing. Accumulation of experience
thus changes the consumer’s relative importance of product quality and price, and spurs the demand
for advanced products. However, part of the consumer’s experience is knowledge on operating a
specific camera, and cannot be utilized after she switches to another one. The consumer thus faces
4 The
latter is inferred from changes in the identity of the cameras.
3
a key tradeoff. On the one hand, learning by doing encourages her to delay switching to advanced
products, since higher human capital brings higher immediate benefit for using the product. On the
other hand, the longer she waits, the more effort she spends on learning non-transferable, cameraspecific features; and because of this, the consumer would rather switch to advanced products
early.
To ensure that alternative explanations are controlled for, I allow for differences across consumers, in their (time-invariant) preferences as well as the way that past history affects their current
choices, which captures across-consumer differences in demand and demand evolution. In addition, the initial period differences, both in prior experience and in the choice of the first camera,
are also captured by the model. Finally, I also model technology evolution, and the individual’s
rational expectation on it.
Both the data and the structural estimates find substantial returns to experience in photography: on average, a consumer with 3 years of experience values her consumption utility with any
given camera more than twice as much as when she just started photography. In addition, not
all the experience gained is generalizable to other cameras: for example, for an average Canon
compact camera user with 3 years of experience, only 2 years of her experience is applicable to
a Nikon DSLR camera. So she faces a switching cost of 1 year of experience, should she switch
to the latter. The switching cost is less pronounced if she would have switched within Canon, and
hence, the lack of applicability of camera-specific knowledge across brands plays a crucial role in
encouraging consumers to be brand loyal.
To my knowledge, this is the first paper to structurally quantify the role of accumulation
in product usage experience on consumer demand. Although the empirical exercise focuses on
choices between entry-level and advanced digital cameras, the insight from this paper can be applied to a broad range of industries, such as home electronics, sports equipment, entertainment,
and other categories where usage of products requires consumer human capital. In studying consumer human capital evolution, this paper contributes to practical understanding of the evolution
of consumer demand through product usage, as well as consumers’ gradual tendency to be locked
in to products with similar characteristics – such as brands.
As the first contribution, this paper quantifies the returns to experience on consumers’ demand
for advanced products. Experience accumulation increases a consumer’s payoff from product us4
age, which in turn increases her demand for advanced products. I find that 1 extra year of human
capital can increase a starting consumer’s demand for the advanced products in the same brand by
26%. This implies that the overall experience stock among consumers has a substantial influence
on the market demand for advanced products, and thus offers an explanation of the demand-driven
innovation hypothesis (Adner and Levinthal, 2001). Supply-side provision of consumer knowledge
– such as free product training, stimulating consumer content creation, or designing products that
are easy to use – can facilitate the evolution of demand via the increase in consumer human capital.
In addition, the consumer herself values the 1-year human capital increase by 405 dollars – which
is equivalent to her discounted total lifetime spending in the entire digital camera market. This
number is in line with casual observations of market price for a photography course, and implies
considerable consumer demand for supply-side provision of knowledge.
Part of the consumer experience is product-specific. As the second contribution, this paper
finds that an important barrier to knowledge transfer is the differences in product designs across
brands, which creates significant brand loyalty that accumulates through experience. As a consumer’s product experience accumulates, she becomes less willing to switch to other brands. I
find that an experienced Canon compact camera user would have been twice as willing to upgrade
to Nikon DSLRs, if her experience were fully applicable to Nikon cameras. Conversely, because
switching across brands becomes increasingly costly over time, a consumer will tend to pick a
brand with favorable long-run characteristics, and attempt to stick to the brand in the future. This
can make products within a brand intertemporal complements – in the sense that a permanent price
drop can increase the sales of other products under the same brand.5 The mechanism for nontransferable human capital is not well-known to the brand loyalty literature, and this paper thus
offers an alternative explanation to brand loyalty (Guadagni and Little, 1983; Dubé et al., 2010),
evolution of consumer brand preferences through experience accumulation (Erdem and Keane,
1996; Bronnenberg et al., 2012), and umbrella branding (Wernerfelt, 1988).
The remainder of the paper is structured as follows. Section 2 gives a brief review to the
literature related to this study. Section 3 describes the data collection process and how I define the
key variables – in particular, the identification strategy that allows us to measure picture quality.
5 That
is, even for mildly forward-looking consumers. The results are produced under an annual discount factor
of 0.54, in line with field estimates from consumer choice data, but far below the commonly used 0.95, which implies
the market interest rate.
5
Section 4 then presents model-free evidence that shows the importance of consumer learning by
doing, and the role of switching cost. Given the evidence, Section 5 outlines an empirical model
of experience evolution and consumer choices on purchasing and using cameras. Next, Section 6
presents and discusses parameter estimates, implied state evolution and price elasticities. Section
7 then discusses the managerial implications, and Section 8 concludes.
2
Related Literature
This paper can be positioned in the intersection of two literatures. On the one hand, my discussion of learning by doing draws from previous theoretical work on consumer human capital
(Becker, 1965; Michael, 1973; Alba and Hutchinson, 1987; Jovanovic and Nyarko, 1996; Ratchford, 2001). Built from the framework in Becker (1965), Michael (1973) and Ratchford (2001)
point out that consumer human capital determines their utility from product consumption. With
different methodology, Alba and Hutchinson (1987) categorize the dimensions of consumer “expertise”, and point out its difference from a consumer’s information set. Jovanovic and Nyarko
(1996) build a framework where non-forward-looking, Bayesian individuals update their knowledge on product usage from previous usage experience, and this increases their incentives to ascend to higher-quality products. Their framework is applied in Foster and Rosenzweig (1995)
in their empirical study of increasing rural labor productivity and the choice of applying a new
agricultural technology.6 Ratchford (2001) constructs a framework for consumer human capital,
and points out its implication for life-cycle consumption, brand loyalty (in particular, related to
its non-transferability) and the decisions to search. Built on this literature, this paper is the first
empirical study using field data to study the effect of consumers’ human capital on their product
replacement/upgrade decisions.
On the other hand, the consumer demand framework of this paper is derived from the literature on dynamic discrete choice of differentiated products, for example, Melnikov (2000), Song
and Chintagunta (2003) and Gowrisankaran and Rysman (2012). In Melnikov (2000) and Song
and Chintagunta (2003), since their interest focuses on first-time adoption decisions, they assume
6A
previous version of this paper also uses the Jovanovic and Nyarko framework, i.e. to specify a Bayesian
updating process for human capital accumulation – human capital as one minus the posterior variance. Applying to
this context, their framework produces similar quantitative insights and a good model fit.
6
away repeated purchases, and hence greatly simplify computation. Gowrisankaran and Rysman
(2012) allow for repeated purchases, but impose a dimensionality-reduction assumption on the
state space, to ease the computational burden. In my paper, the focus is on re-purchase rather
than first-time purchase decisions, and I need to consider endogenous product usage decisions and
the corresponding outcome – in this case the picture quality. I also take into account dynamic
optimization under differentiated product characteristics. Specifically, to maintain the key feature of evolving consumer human capital as well as accounting for other (high-dimensional) state
variables, I impose a dimensionality-reduction assumption that is in spirit of Gowrisankaran and
Rysman (2012), but does not require the extra layer in the fixed point algorithm.
The fact that consumer human capital is not perfectly transferable creates a switching cost. The
general topic of switching cost relates to the empirical literature on the effect of switching costs on
consumer decisions, in grocery shopping (Dubé et al., 2010), pharmaceutical products (Crawford
and Shum, 2005), health care (Nosal, 2012), and many other categories. In this literature, there
are various explanations to a consumer’s lack of willingness to transition across brands – hence
“brand loyalty”. In this paper, I propose an alternative mechanism: that consumers are brand loyal
because it is difficult for their experience to transfer across brands – possibly due to differences
in designs. A similar explanation, “skill-based habits”, is proposed by Murray and Häubl (2005)
in their experimental studies. In addition, I also demonstrate that this has dynamic implications
especially for forward-looking consumers.
This paper is also related to the empirical literature on the effect of consumer learning. This
literature (Erdem and Keane 1996; Erdem et al. 2005, among others) characterizes the effect of
information of product attributes on consumer demand. In this framework, knowledge also endogenously evolves through past purchase experience, but the main effect of such knowledge is on
consumers’ belief (i.e. their information sets), while in my model, experience is effective on consumers’ ex post utility from product usage.7 Empirically speaking, learning on product attributes
tends to stop rather quickly,8 while in the case of learning by doing, I examine changes in consumer
7 The difference also corresponds to the difference in “familiarity” and “expertise” in Alba and Hutchinson (1987).
In Nelson (1970), the different explanations are two aspects of his categorization of experience goods: “After using
..., its price and quality can be combined to give us posterior estimates of the utility of its purchase.” [Nelson (1970),
“Information and Consumer Behavior”, p.313].
8 For example, Dubé et al. (2010) do not find non-stationarity in the choice pattern for products that are not new to
the market.
7
choice patterns over the course of up to 10 years.
3
Data
3.1
Collection
I extract picture level data from Flickr.com – a popular photo sharing website. Flickr started its
business in 2000 by Ludicorp, and was acquired by Yahoo! in 2005. The data extraction was
implemented between March 2012 and April 2013, until a major change in user-interface took
place on Flickr. During the data collection period, pictures (including their detailed information)
were publicly viewable, even without a user account.
Camera-recorded information is embedded in each picture, as Exif (exchangeable image file
format) data. For the purpose of this paper, those data contain valuable information for camera
identity, as well as the date of capture. To complement the Exif data, I also collect information
on the date of upload, and the cumulative views and “favorite” votes from the upload till the dataextraction time. Figure 1 summarizes the information I get from each picture.
I collect data at two levels. At the picture level, I sort an individual’s pictures in order of upload
dates, and collect the data once, from one in every five pictures. This gives me cross-sectional data
on picture level information.9 At the individual level, I collect data on Flickr-summarized monthly
picture-taking and uploading records, for each individual.
I also gather a cross-sectional data-set for camera characteristics, and a longitudinal auction
price data-set. The camera characteristics data-set is compiled from the Flickr camera database,
DPreview.com, and Cnet.com. In addition, Pixel-peeper.com summarized a long monthly price
history for average Ebay auction prices per camera, from 2006 to 2013.
9I
did not re-visit a picture multiple times, because the time spent on collecting data from each picture is large.
8
Figure 1: Structure of picture-level data from Flickr
Note: This figure depicts the structure of my picture level data, extracted from Flickr.com. The vertical dashed line
divides data that are originally recorded by the cameras (Exif data, embedded in each picture), and data that are
originally recorded by Flickr. From the camera-recorded data, I collect the camera identity (in this example, Nikon
D60) and capture date. From Flickr-recorded data, I collect the upload date, as well as the cumulative views and
“favorite” votes from upload to data-extraction. This is done once per picture.
9
Table 1: Sample Selection Criteria
percentage
Taken by compact camera or DSLR
96.4
Exif data complete
89.6
Taken after year 2000
89.0
All above criteria
75.6
obs.
2777753
Notes: This table reports the sample selection criteria. On the picture level data, I drop the observations on pictures
taken by camera formats other than compact camera or DSLR; or with Exif data that lack an indicator of camera model
or date of capture; or taken no later than January 1, 2000. Altogether, this excludes 24.4% of the sample.
3.2
Sample selection and summary statistics
I focus on the “pro account” (paid account) users, who are able to upload many more pictures
than the free account users.10 The free account users tend to upload few pictures even over many
years. Hence, one might run into the risk of omitting certain camera-switching decisions. On
the other hand, however, pro accounts are less representative for the entire market. This paper
focuses on within-individual changes, and hence does not require data to be representative over
demographics. However, the quantitative conclusions from the counterfactual experiments should
not be over-generalized.
I collect data from all paid account users with a user-name no long than 5 letters/digits. Focusing on shorter user-names gives me users with long histories on Flickr, and usually enables us
to observe camera usage in long time spans.11 On the other hand, it is reasonable to assume that
user-names are exogenous to the variables of interest. Sampling one in every 5 pictures gives me
close to 2.8 million observations on the picture level. Among these data, I disregard the pictures
taken by cell phones, film cameras, camcorders or digital media players, or those claimed to be
taken prior to year 2000 (which is more likely to be a mistake in the camera date settings), or have
incomplete Exif data (in particular when identities of the cameras or the picture taking time are
missing). This excludes 24.4% of the picture data – as shown in Table 1.
Table 2 provides summary statistics for the user level data after sample selection. There are
10 Flickr offers either a free account – which is imposed a monthly upload capacity as well as a maximum-viewablepictures restriction, or a “pro account” that costs $24.95 (as in 2012) annually.
11 The underlying assumption is that the naming strategy is orthogonal to preference and experience. On the other
hand, the in-sample duration is not orthogonal to preferences – and hence I do not select on it.
10
Table 2: User Level Data Summary
Mean Median
months since registered in Flickr
69
74
number of contacts at data extraction
94
20
total number of pictures
1691
981
number of in-sample pictures
359
203
number of cameras ever used in-sample
4
3
max views per month, first 10 pic
7
1
max views per month, last 10 pic
20
4
price of the least expensive camera used 216
157
price of the most expensive camera used 1040
762
obs.
5499
5499
Stdev
24
292
1897
410
4
57
193
191
655
5499
Notes: The table reports summary statistics from the data. Mean, Median and StDev are the mean, median and
standard deviation of the data, respectively. The number of contact is the number of other accounts, who are followed
(subscribed) by the given user at the time of data extraction. The number of cameras ever used in-sample is the number
of unique camera identities one observes from the user’s Exif data. Prices of the least and most expensive cameras are
in 2005 US dollars.
three interesting points to note. First, the median duration of observation for a user is beyond 6
years. This is a long enough period to observe the slow evolution of an individual’s photography
knowledge. Second, among the 6 years in Flickr, the median user only subscribed to 20 other users.
Compared to Facebook users, this shows that (this sample of) Flickr users are not social-network
driven.12 Second, there is a considerable increase in the maximum views per unit time among
pictures taken at the beginning of the sample, compared to those taken at the end of the sample;
while the views have a larger spread towards the end of the sample. This suggests both an increase
and a divergence in the number of views one’s pictures can attract.13 Third, the median individual
has had 3 cameras throughout the 6 years’ in-sample period, while there is considerable dispersion
in the prices of her camera: the real (Ebay auction) price of her most expensive camera is more
than twice of the price of her least expensive camera.
12 As a comparison, the median Facebook user has 200 friends, by account of Aaron Smith (extracted in June, 2014,
from http://www.pewresearch.org/fact-tank/2014/02/03/6-new-facts-about-facebook/).
13 Which might be due to changes in picture quality, or changes in the size of user base of Flickr.com.
11
3.3
Picture quality
In this section, I construct an index of picture quality, which will be later treated as data in the
reduced form analysis and structural estimation. The basic idea is that a large amount of views of
a picture might either suggest that it has high quality, or that a period of high popularity of Flickr
coincided with the period when the picture was on display. I exploit the variation between the
date of capture of a picture, and the date when it was uploaded. Holding the date of upload fixed,
differences in views among the pictures should solely reflect differences in their quality – as we are
effectively holding the flow of viewers to be the same. My sample consists of more than 158,000
user-months of upload combinations. Among those uploaded in the same month, the first picture
was captured 4 months earlier, on average, than the last picture. This gives me ample variation in
the capture dates to measure picture quality.
Formally, I model the cumulative number of views of picture p captured by individual i, as
the accumulation of an underlying viewer-flow process to the photographer i, f lowipt , which is by
itself multiplicative in the quality of the picture qip , the overall flow of viewers into Flickr.com φt ,
and other observed characteristics of the picture that are not related to quality, zip (e.g. the order
that pictures are displayed might affect their views):
viewsip =
∑
f lowipt
(1)
t0 ≤t≤t1
where
f lowipt = φt exp (qip + zip ψ) .
Omitting i and p subscripts, I denote t0 and t1 to be the calendar dates of upload and data extraction,
respectively. Note that t0 and t1 are picture specific. The cumulative number of views is the
summation of the viewer flow between these two dates. In the viewer flow specification, qip is the
(unobserved) quality of the picture, which is implicitly a function of user experience, camera, and
an econometric error.14
14 One
might alternatively interpret this as a noisy measure of picture quality.
12
Take the log of Equation (1), we have
log (viewsip ) = Φt0t1 + zip ψ + qip ,
(2)
noting that Φt0t1 = log ∑t0 ≤t≤t1 φt is a time-window-specific fixed effect, that captures the overall
cumulative viewer arrival in the time window [t0 ,t1 ], when the picture was on display.
To measure picture quality, I estimate Equation (2) by ordinary least squares, controlling for
combinations of picture upload month and data extraction month (Φt0t1 ), as well as individual fixed
effects (contained in qip ). I also include the following control variables in zip : 1) the topic of the
pictures, as captured by tag fixed effects, 2) the number of pictures uploaded in the same batch,
3) the order of the focal picture in the upload batch, and 4) months since a user was registered
on Flickr (as a proxy of the accumulation of friends networks). All control variables are coded as
dummies, allowing the specification to be as flexible as possible. I take the projected individual
fixed effect plus the residual term, as a proxy of picture quality.15
Table 3 summarizes the maximum picture quality q̂ip in each picture-taking month, which
characterizes the quality of pictures that an individual can produce. One can immediately spot the
following patterns.
First, with accumulating years of experience, the individual can produce increasingly higher
picture quality, up to a point where knowledge has been saturated, and the change in picture quality
is statistically negligible. In other words, there is a clear pattern of learning with decreasing speed.
Second, using a small subsample with non-zero favorite-votes data,16 one can cross-check
whether the developed measure of picture quality is reasonable. I find that the correlation between
maximum picture quality and maximum rating (if nonzero) is around 60%, which justifies that the
maximum quality is a reasonable measure of the outcome of picture taking.17
15 The reason I consider individual fixed effects as systematic across-individual difference in quality rather than
other factors such as being popular on Flickr, is because we can trace every user to her starting point in Flickr,
but not to her initial experience in photography. Therefore, it is much more plausible to think of heterogeneity in
initial conditions as heterogeneity in skills. As a robustness check, leaving out the individual fixed effect does not
qualitatively change the (reduced form and structural) estimates.
16 The share of individual-monthly observations where at least one picture has received at least one favorite vote is
15%.
17 As a side note, the mean quality has a poor correlation with the average rating: around 10%.
13
Table 3: Summary of the monthly highest inferred quality
max quality stdev max favs corr. with qual.
obs
0 year of expr
0.619
1.503
0.636
0.583
3704.000
1 year
1.537
1.747
0.938
0.625
2561.000
2 years
1.835
1.852
1.034
0.648
2788.000
3 years
2.054
1.832
1.073
0.628
3026.000
4 years
2.091
1.782
1.087
0.612
3240.000
5 years
2.045
1.731
1.064
0.600
3290.000
Notes: This table summarizes the individual-monthly maximum of the inferred picture quality (Section 3.3), which
is treated as data in the subsequent analysis. Monthly maximum refers to quality of the best picture captured in the
given month, by an individual. Years of experience is defined as number of years from the first in-sample picture to the
current month of picture-taking. The first two columns summarize its mean and standard deviation. The third column
presents average of the highest rating (“favorites”) one gets for pictures taken in the month, given that the highest
rating is non-zero (15% of the individual-month data). The fourth column presents its correlation coefficient with the
highest inferred quality. Finally, the fifth column shows number of observations on the individual-month of capture
level.
3.4
Camera ownership
I next infer camera ownership from the Exif data behind each picture. As previously mentioned,
the camera identity is embedded in the picture’s Exif data. I then assume that, for a given individual, whenever one observes a new camera capturing its first in-sample picture, I assume that the
previous camera has been replaced.
For 75% of all individual-camera combinations, I never observe an old camera taking pictures
after the arrival of a new camera. For the remaining 25%, although the earlier cameras still take
pictures, the majority of the pictures are taken by the most recent cameras acquired. To document
this, Figure 2 presents the joint probability of the number of cameras that are still active (i.e.
cameras that would take pictures later than this date), and the occasions that the newest cameras
taking most pictures. Overall, there are few cases when the latest camera is not the most active
one.
14
.8
0
.2
probability
.4
.6
latest camera takes majority of pictures
more pictures taken by other cameras
1
2
3
4
5
Figure 2: Joint probability of the number of cameras owned, and whether the latest camera takes
the most pictures (x-axis: # cameras)
Note: This figure shows the joint probability of the number of cameras owned at a given time, and the incidence that
most pictures are taken by the latest camera. A camera is owned at a point in time if I observe at least one picture
taken before that, and at least one picture taken afterward. By construction, a camera takes all pictures if it is the only
“owned” camera – as represented by the dark bar at x axis = 1. When more than one camera is present, I find that in
most occasions, the last camera still takes most pictures.
3.5
Computing price indices
For digital cameras, as other consumer electronics, prices vary a lot among retailers, fluctuate over
short periods of time, and display large differences across first- and second- hand markets. I cannot
observe the actual prices that the consumers observe. Instead, I observe the monthly average Ebay
auction price for each camera model. This price data are averaged across first and secondary
markets. Especially for older camera models, it better represents the prices the consumers face
compared to a retailer’s list price.
With this data, I first deflate prices to 2005 US dollars. Then, separately for both camera
formats, I take the weighted average of the prices of all available cameras in a given month, by
their market shares in the Ebay auction data.1819 Since the data only ranges from 2006 onward, I
interpolate the missing values before 2006, by taking a log-linear fit against time, plus a simulated
18 That is, the number of auctions for a given camera model, as a percentage of the total number of auctions in the
sample.
19 In this version, I do not consider other camera characteristics such as resolution. In robustness-check versions
when this was considered, I use the same method to compute resolution indices.
15
regression error.20
4
Reduced form evidence
4.1
Overview
This section first presents descriptive evidence on an average individual’s increasing tendency to
use advanced cameras, even when calendar time effects are controlled for. Then, I present the
following reduced-form evidence that point to consumer learning by doing: 1) complementarity
between a consumer’s experience and her usage of an advanced product; 2) the imperfect transferability of consumer experience upon camera switching; and 3) the long-run evolution of picture
quality, and the increasing attrition of knowledge if the consumer switches cameras.
4.2
Increasing tendency to using the advanced products
I first screen out individuals who I observe for less than 5 years, and only look at the remainder of
the sample in their first 5 years. This avoids selective attrition – in this case, that the individuals
with lower ability to generate picture quality might systematically drop out earlier. With this
screening criterion, we have the same set of individuals in every time period for the reduced form
evidence.
I first present evidence that suggests an increasing tendency to use advanced products, as one’s
experience in photography grows. To pin down experience effects, I control for the role of technology and other sources of calendar-time effects, by estimating a linear probability model with
the choice of camera format on the left-hand side, and experience dummies and calendar time
dummies on the right-hand side:
60
2013
Formatit = αi + ∑ βexpr,t (exprit = t) +
t=1
∑
βyear,y (yearit = y) + εit .
(3)
y=2001
Essentially, controlling for the calendar time effects allows us to compare within a given point in
20 Separately for each format, I regress log price index on a linear time trend, and interpolate the missing value using
the linear prediction plus a simulated prediction error. The R-squared for the linear regression are around 0.7 for both
formats. Keane and Wolpin (1994) use this method to interpolate missing data in their value function calculations.
16
.4
share of DSLR cameras
.45
.5
.55
share of DSLRs
90% CI
0
1
2
3
years (0 = first picture)
4
5
Figure 3: Product-format choice and user experience (selected sample)
Notes: This figure depicts the changes in the choice of product format (DSLR vs compact camera), given a user’s
years of experience – defined as the number of years since one’s first in-sample picture. It also depicts (part of) the
90% confidence intervals. To control for selective attrition, I choose the subset of individuals whom I observe for
no less than 5 years, and only focus on their first 5 years of data. To control for advances in technology (and other
calendar-time effects), I estimate a linear probability model in Equation (3), and present the estimated α̂i + β̂expr,t as
the calendar-time-detrended estimates in camera usage choice.
the calendar time, across individuals with different experience stock at this point.
The resulting tendency to use DSLR cameras – calendar-time de-trended – is shown in Figure
3. Despite the large confidence interval (which might be due to the amount of control variables
and fixed effects), I find that having 5 years of photography experience raises one’s tendency to use
an advanced camera by 10%. This suggests that experienced consumers are more likely to choose
advanced products, given a constant technology level.
4.3
Complementarity between advanced cameras and human capital
The next 3 pieces of evidence picture an explanation to the increasing shares for advanced products.
As the second piece of evidence, I show that there is complementarity between using an advanced
product (i.e. a DSLR) and having rich experience in picture taking, in producing high-quality
pictures. To show this, I take the within-individual-month difference between the highest picture
17
.4
gains from advanced cameras
0
.2
−.2
0
1
2
3
years (0 = first picture)
4
5
Figure 4: Complementarity between experience and advanced products (selected sample)
Notes: This figure shows complementarity between experience and advanced cameras, in the sense of increasing each
others’ productivity in generating maximum monthly picture quality. Within individual-month, I take the difference
between monthly highest picture qualities generated by DSLR and compact cameras, for those who used both in the
month (6.9% of the individual-month observations). This is the “gain from DSLR” measure in the Y axis. The solid
line is local polynomial fit against experience, while dashed lines are 95% confidence intervals. To eliminate selective
attrition, I limit the sample to consumers we observe for more than 5 years.
quality generated by a DSLR camera and a compact camera – for a small subsample of individuals
using both formats in a given month. This measure, “gain from DSLR”, is defined in 6.9% of all
individual-month observations.
Plotted against experience, Figure 4 find that a starting consumer does not find that using a
DSLR contributes to her ability to produce high-quality pictures. This quickly changes when
she gains the first year of experience, as the consumer now finds a 0.2-unit difference in picture
quality, between using a DSLR and a compact camera. This provides within-individual evidence
on the complementarity between knowledge and product quality. Also note that the difference
starts slightly above zero, which indicates that even new consumers do not find a DSLR camera
generating lower picture quality than a compact camera.
Alternative between-individual evidence, presented in Appendix Figure 12, is qualitatively
similar.
18
4.4
Changes of picture quality at camera-switch
Thirdly, I present evidence for the imperfect transferability of consumers’ knowledge on camera
usage – or human capital, across different cameras. To do so, I normalize the date of cameraswitching to be period 0, and look at the highest quality an individual produces in a given month,
around the time when she switches products. Figure 5 shows that there is an immediate drop in
picture quality at switching. The drop in picture quality indicates that not all the knowledge from
the previous product is transferred to the new camera.
Note that the highest picture quality drops despite that the individual produces more pictures
immediately after switching – which would have generated higher picture quality if there were no
human capital attrition. Also note that the pattern is robust (besides being noisier) when conditioning on the direction of camera switching, as presented in Appendix Figure 11.
In addition, after the camera switch, the picture quality quickly goes up in the first 3-4 months,
and it further gradually increases to a higher level in 1.5 years after the switch. This suggests
that at the instance of camera switching, the individual loses both explicit knowledge on camera
operations (e.g. menu and button layout), as well as implicit knowledge on camera usage (e.g. how
to best circumvent a certain product limitation). While the first can be quickly learned in a month
or two, the second can only be learned with long experience with the new camera.
4.5
Evolution of picture quality and switching cost
Finally, in separate panels, Figure 6 presents the evolution of consumers’ monthly maximum picture quality, and the decrease in quality when using a new camera, conditional on her experience.
In the left panel, monthly maximum picture quality increases sharply at early experience levels,
and gradually stops increasing from year 3 onwards. This shows that consumer learning by doing
enhances their human capital on camera usage, but the learning speed is lower when one already
processes knowledge. The figure can be plotted separately for compact camera and DSLR users.
Shown in Appendix Figure 12, the concave trend persists, while DSLRs produce significantly
better picture quality than compact cameras.
In the right panel, I contrast the maximum picture quality produced by consumers who have
used their cameras for more than 3 months, and those who have been with their camera for no
19
10
8
9
number of pictures taken
1.3
max quality in a month
1.1
1.2
1
−1
0
1
years (0 = camera switching)
7
.9
max picture quality
number of pictures taken
2
Figure 5: Quality of best pictures around camera switching
Notes: This figure depicts the changes in maximum picture quality, around the period when an individual switches
cameras. We focus on the years before and after a consumer switches her camera at year 0. With the left vertical axis,
the dark line (and dashed lines as its 95% confidence interval) depicts the maximum picture quality that the consumer
can produce, using her old camera until year -1/12, and new camera from year 0. The dash-dot gray line depicts the
number of pictures that the consumer takes in each month. Both lines are local polynomial estimates with bandwidth
1.
20
differences in quality
.6
−.2
.8
max picture quality in a month
1
1.2
diff in quality between familiar and new cameras
−.1
0
.1
.2
.3
.4
.5
1.4
max picture quality
0
1
2
3
experience (years)
4
5
0
1
2
3
experience (years)
4
5
Figure 6: Quality of best pictures over photographic experience
Notes: Left panel: monthly highest picture quality for a given individual, against (general) experience in years. Right
panel: the difference between monthly highest picture quality produced by individuals who are using cameras that have
been purchased more than 3 months ago, and those by individuals who are using cameras purchased more recently,
plotted against years of experience. For both panels, the solid circles are mean values across individuals, while brackets
are 95% confidence intervals. To eliminate selective attrition, I limit the sample to consumers we observe for more
than 5 years.
more than 3 months. Users of new cameras systematically generate lower picture quality than
users who are familar with theirs, which agrees with the drop in picture quality at switching,
in Figure 5.21 In addition, when plotted against years of (general) experience,22 I find that at
the start of the sample, the picture quality generated by an inexperienced consumer shows no
statistically significant difference between new cameras, and cameras that the consumer is familiar
with. On the other hand, for those with longer (general) experience, the switching cost is significant
and increasingly important. This suggests that as the consumer learns, a share of the incoming
knowledge is product-specific – the accumulation of which generates the increasing switching cost
profile.
21 The picture quality generated by new cameras, and cameras that the consumers are familiar with, can be found
in Appendix Figure 13.
22 This is defined as the duration between the production of the first in-sample picture, and the current month of
picture taking.
21
5
Structural Model
5.1
Overview
This section presents the structural empirical model. Whereas the model on durable good purchase
with learning by doing is general, it will be presented in the context of digital camera markets for
concreteness. A bare-bone version of the model with numerical results will be presented in Section
B in the Appendix, to highlight the key properties of a demand model with (general and specific)
learning by doing.
In the model, I jointly characterize a consumer’s decisions to purchase digital cameras, and
her decisions to use the product. Combining a camera and the stock of experience – or “human
capital” – produces pictures that generate consumption utility,23 and at the same time, contributes
to the consumer’s human capital stock. Therefore, past usage decisions build up consumer human
capital, and hence future utility. With rational expectations and a non-zero discount factor, the
consumer makes camera replacement and usage decisions, taking into account the consequences
of her decisions on her future human capital stock.
5.2
Decisions on camera replacement and usage
Consumer i in each period t = 1, ..., T decides whether to purchase a new camera, and in the case
of purchase, which format and brand to buy. If the consumer buys a new camera, she replaces the
old one with no resale value. Afterward, she decides whether or not to take pictures in this period,
and if she does so, she derives utility from the highest picture quality generated by her camera and
her stock of human capital.
I denote the decision as Ait = (Bit , Dit ), where symbols A, B, and D stand for “action”, “buy”
and “do”, respectively. The discrete variable Bit > 0 denotes the choice of buying a new camera,
in which case the consumer chooses among combinations of one of the two formats (a compact
camera or a DSLR) and one of the three brands (Canon, Nikon and “other brands”). In the case of
no purchase, I denote Bit = 0.
23 I
follow the terminology in Michael (1973) and Foster and Rosenzweig (1995). Alternative terminology include
“know-how” (Besanko et al., 2010), and “expertise” in Alba and Hutchinson (1987). I also use the term “knowledge”
interchangeably with human capital, and this is not to be confused with information.
22
State variable Kit = 1, ..., 6 denotes the identity – the brand-format combination – of the camera
that the individual owns at the end of period t. If a camera is purchased, the consumer replaces the
previous camera that she owned with the new one, i.e.
Kit =


Kit−1
if Bit = 0

Bit
if Bit > 0.
(4)
I do not consider resale, or multiple camera ownership.
The binary variable Dit denotes the decision of whether to take pictures (Dit = 1) or not (Dit =
0), using the latest camera Kit , i.e. after the replacement decision. If she decides to use the camera,
she incurs a cost of effort ei , which summarizes the dis-utility or utility from taking pictures in a
period. Also, she takes one draw that determines the realization of the highest picture quality –
denoted Qikt – from which she derives her consumption utility.
To keep the model simple, I do not model the decision on the number of pictures to take.
Modeling this aspect will also necessitate modeling of picture selection and upload decisions,
which are not central to the core mechanism.
The timing of consumer decisions and evolution of the state variables are graphically presented
in Figure 7.
5.3
Learning by doing
Learning by doing is reflected by the accumulation of the stock of human capital, or the productive
knowledge that the consumer has on her camera. I model human capital accumulation through
pictures taking, in a way that is similar to the labor/firm learning by doing literature (Besanko
et al., 2010, 2014; Shaw and Lazear, 2008; Levitt et al., 2013). Human capital Hikt on camera k
accumulates by one unit at the end of the period, whenever the consumer has taken pictures.
Hikt+1 = Hikt + Dit .
23
(5)
Figure 7: Timing of decisions and evolution of the state variables
Notes: The figure presents the timing assumptions of consumer decisions and state evolution, in a given period.
As normalization, the individual has 1 unit of initial human capital at the start of the sample:24
Hi1 = 1.
5.4
Switching cost
Human capital is product specific, and hence cannot be fully transferred to other cameras. For
example, knowledge on menu layouts of one camera cannot be fully applied to the others. Hence,
there is attrition on consumer human capital when she switches from camera k to k0 . Specifically,
motivated by Figure 6, I impose that the switching cost is proportional to the current human capital
stock:
Hikt − Hik0t = sikk0 · Hikt ,
(6)
where the switching cost sikk0 is set to be symmetric in k and k0 . Further restrictions will be imposed
on sikk0 to reduce the number of parameters to be estimated.
24 I
normalize it to 1 so that the productivity of picture quality is increasing in the shape parameter κ, introduced in
Equation (8). Heterogeneity in the initial experience is captured by individual-specific intercept in Equation (7).
24
5.5
Production function
The individual derives utility from the quality of the best picture she produces. This is the output of
a production function of the “physical capital quality” – a common function f of the brand-format
combination of the current camera – and an individual-specific, concave function of the human
capital stock, gi . Formally,
Qikt (Kit ,Yit , Hikt ) = qi + f (Kit ,Yit ) · gi (Hikt ) + ηikt
(7)
and
6
f (Kit ,Yit ) =
∑ γk 1 (Kit = k) +Yit ,
k=1
where Yit = ∑2013
y=2001 φy 1 (yeark = y) is the year-of-introduction effect of camera model k on picture quality,25 and I name it “technology index”. Given the separable structure, γk captures timeinvariant camera format effect, whereas φy captures the time effect, or a general trend of technology, in the quality of pictures. Overall, the function f (Kit ,Yit ) captures the camera specific returns
to human capital stock.
On the other hand, I restrict
κi
gi (Hikt ) = Hikt
,
(8)
where κi ∈ (0, 1) dictates the concavity of picture quality with respect to the human capital stock.
In the previous version of this paper, I presented a model with a non-parametric learning curve,
which generates qualitatively the same results.
Finally, ηikt is an independent and identically distributed (IID) econometric error, that captures
non-systematic variation in the maximum picture quality.26
5.6
Utility
In the model, the consumer derives per-period utility from 1) consuming the quality of the best
picture she made in this period, 2) effort spent taking pictures, 3) her expenditure on buying new
25 The
i,t subscript in Yit captures the identity of the camera model (beyond the brand-format combination).
alternative way to capture the effect of technology is to introduce observed characteristics – the most popular
one being camera resolution. This is implemented in earlier versions, and I found that resolution has little explanatory
power on the measured picture quality.
26 An
25
cameras, and 4) the immediate satisfaction from buying new cameras. Considering these four
aspects, and denote I model the utility function as
ũiat = uia (Kit−1 ,Yit , Hik0t , Pk0t ) + εiat
= (αi · E [Qik0t |Kit ,Yit , Hik0t ] + ei ) · 1 (Dit = 1) +
∑
0
βi1 Pk0t + βi2 Pk20t · 1 Bit = k0 +
k 6=0
∑
∑ λi,kk0 1
0
0
Bit = k ,Cit−1 = k + εiat .
(9)
k 6=0 k
where, following notational conventions in the above section, I denote k0 as the realization of Kit ,
and k as that of Kit−1 . These four sources of utility are explained below.
1) At the time when making decisions, the individual derives linear utility from the expected
picture quality, since the idiosyncratic shock on the maximum picture quality, ηikt , is not yet realized at the point of decision. I allow the marginal utility on picture quality to be heterogeneous
across individuals. Denote it αi .
2) The consumer takes pictures in some periods but not in others. I allow for a “utility of
effort” – parameter ei – to incur whenever pictures are taken in a given period. This captures the
(dis-)utility from taking pictures, if one were to generate zero quality.
3) To control for systematic variation in the camera purchase decisions that are not explained
by the evolution of human capital, I consider the conventional price effects. Specifically, I impose
quadratic dis-utility from price spent, when the individual purchases camera k0 (i.e., Bit = k0 > 0).
In this case where price dispersions are large, the quadratic functional form captures that marginal
dis-utilities from spending an extra dollar might be different on a 80-dollar compact camera, and
on a 800-dollar DSLR.27
4) To account for other variation in the choice patterns, I allow for immediate utility impacts
from purchasing new cameras, which can be history-dependent. Further restrictions are placed in
Section 5.10.
27 A
linear specification will not fundamentally change estimates of the other parameters, but will predict very
different elasticities for DSLRs and for compact cameras, while a natural log specification will overly flatten the
utility profile, within common price range for DSLRs.
26
5.7
State space and transitions of the state variables
I denote the state variables as Siat = (Kit−1 , Yit , Hikt , Pt ). Here, Kit−1 is the brand-format combi
nation of the camera at the beginning of the period. Yit = Yit−1 , Ỹt – the “technology index” –
summarizes the technology level of the previous camera (Yit−1 ) as well as the current market “stateof-the-art” technology (Ỹt ). Hikt is consumer human capital with respect to the camera Kit−1 = k
at the beginning of the period. And finally, Pt is a vector of price indices for all brand-format
combinations on the market at time t.
The transition processes for camera brand-format Kit−1 and human capital Hikt are characterized by the camera replacement Equation (4) and the human capital transition Equations (learning
by doing (5) and switching cost (6)).
Finally, I assume that individual decisions do not affect the market equilibrium prices and the
market technology level. The individuals are price takers, who rationally expect an exogenous
price-index transition matrix
Π pp0 = Pr pkt = p0 |pkt−1 = p .
In addition, I assume that the “personal” technology Yit stays constant if an individual does
not make a purchase. If she does, however, she expects the technology of the new camera to
immediately jump to the market technology level. On the other hand, the market technology index
Ỹt follows an exogenous Markov process on its own. Specifically,
Yit =


Ỹt
if Bit 6= 0

Yit−1
if Bit = 0.
(10)
and
Ỹt = χ0 + χ1Ỹt−1 + ωit .
(11)
The Markov assumption on the market technology follows the ideas in Gowrisankaran and
Rysman (2012) and Hendel and Nevo (2006),28 in the follow sense. On the one hand, it is a di28 Gowrisankaran
and Rysman (2012) assume that the discounted sum of future utility is Markov, while Hendel and
Nevo (2006) assume that a part of the individual flow utility is Markov.
27
mensionality reduction assumption, so that in computing the optimal decision rules, the researcher
does not have to keep track of each of all the picture quality-relevant state variables, but rather,
a sufficient statistic of them. On the other hand, the assumption conjectures that the individual
does not know exactly what technology level she is going to get in the next period, but her Markov
belief is on average correct.
5.8
Dynamic programming
With rational expectations, the individual makes purchase and usage decisions every period by
maximizing the sum of discounted flow utilities, or solving
max ∑ δ τ−t Et [uia (Siaτ ) + εiaτ ] .
a
τ≥t
Given stationarity assumptions on the function uia (·) (as in (9)) and transition process of Pt , this
is a standard dynamic decision problem in spirit of Rust (1987) and others, where the consumer
solves the equivalent static decision problem
max Uia (Siat ) + εiat
a
where the choice-specific value function Uia (Siat ) is defined by the Bellman equation
h
i
Uia (S) = uia (S) + δ · E max Uia S0 |S, a .
a
5.9
(12)
Identification
Section 3.3 discussed identification of picture quality, from cross-sectional data of picture taking
and posting dates, and their cumulative views. With exogenous variation in the picture-taking
dates, within pictures that are posted at the same time, one can separately identify the popularity
of the Flickr website and of each individual photographer, from the quality of their pictures. Table
3 documents that the inferred picture quality follows “intuitive” patterns, in the sense that it is well
correlated with the ratings data, for a small subsample of pictures with nonzero ratings. Appendix
A performs robustness checks for when the upload decisions are selective.
28
Next, I treat each individual’s monthly maximum picture quality as data, and separately identify the evolution of picture quality, and individuals’ utility from picture quality and other market
characteristics. The production function and the evolution of human capital is identified by the
observed evolution of picture quality, as a function of the previous period quality, initial picture
quality, identity of the camera, and the camera usage and switching decisions. With panel data, systematic variations across individuals detect heterogeneity in the evolution of human capital, while
the production function parameters are the same across consumers. Of course, this is achieved
by controlling for endogenous camera usage and switching decisions, which are modeled in the
dynamic discrete choice problem and jointly estimated.
Identification of the dynamic discrete choice model, treating picture quality as a state variable
with known evolution, is achieved under standard conditions as in Magnac and Thesmar (2002)
and Kasahara and Shimotsu (2009).
5.10
Implementation
5.10.1
Production function parameters
In the production function, I assume that the returns to human capital, γk , are the same for the same
format of cameras. This reduces the computation burden of estimation. As a robustness check,
relaxing this restriction does not introduce noticeable differences among γk0 s between brands.
5.10.2
Switching cost
To further parameterize the switching cost sikk0 , I allow it to vary across the cases when the consumer switches within the same format of products, or across formats, or across brands. I assume
that switching across formats incurs no smaller switching cost than within a format; and similarly,
switching across brands incurs no smaller cost than within a brand. To impose these assumptions,
I specify the following structure for the switching cost across formats and across brands:
1 − sikk0
f ormat
baseline
brand
= 1 − si
· 1 − si
· 1 − si
29
f ormat
where si
and sbrand
symbolize the across-format and across-brand switching cost, taking value
i
0 when the individual switches within format or brand, respectively.29
5.10.3
Choice intercepts and other explanations of state dependence
The utility function in (9) gives a very general specification of choice state dependence and choicespecific intercepts, that does not depend on the potential picture quality one generates. In implementation, I restrict the utility specification to a more parsimonious structure, which is characterized by 5 parameters:
λDSLR 1 (Bit ≥ 4) + λCanon 1 (Bit = 1, 4) + λNikon 1 (Bit = 2, 5)
+λFormatSwitch 1 ( f ormatit 6= f ormatit−1 ) + λBrandSwitch 1 (brandit 6= brandit−1 )
where λDSLR captures the immediate utility of purchasing a DSLR camera (relative to a compact
camera),30 λCanon and λNikon capture the immediate utility of purchasing specific brands, while
λFormatSwitch and λBrandSwitch capture format- and brand- switching effects (in additional to the
switching cost in human capital).
5.10.4
Heterogeneity
To capture heterogeneity in the preferences and the human capital formation processes, I assume
that there exists a finite-mixture of permanent individual heterogeneity, both in the utility function
and in the evolution of human capital. That is, I allow individuals to have different initial picture
quality, different learning speeds, different switching costs, and different utility parameters. To
implement this, I normalize the initial human capital to Hi1 = 1. At the same time, I allow the
production function intercepts qi , shape parameter κi , as well as the switching cost sikk0 , to be
heterogeneous across individuals. On the other hand, I impose that the production function (7) is
homogeneous across individuals. The parameters to estimate are a vector of production function
29 For example, from a Canon compact camera, switching to another Canon compact camera costs 1 −
baseline · 1 − sbrand ; switching to a Canon DSLR
1 − sbaseline
; switchingto a Nikon compact
camera
costs
1
−
1
−
s
i
i
i
costs 1 − 1 − sbaseline
· 1 − sif ormat ; and finally, switching to a Nikon DSLR costs 1 − 1 − sbaseline
· 1 − sif ormat ·
i
i
1 − sbrand
.
i
30 I cannot estimate a separate compact camera utility because the two brand coefficients almost capture the entire
market, so a λCompact and λDSLR together will produce close-to-perfect colinearity with the brand parameters.
30
coefficients γk , ση2 , as well as a vector of heterogeneous utility and human capital evolution
parameters (αm , βm , em , qm , κm , sm ), where m = 1, 2 to represent parameters for one of the two
segments.
5.10.5
Initial conditions
Heterogeneity in the prior-to-sample experience is characterized by the heterogeneous production
function intercept qm .
Choices of the initial cameras are endogenous to preference heterogeneity. To endogenize the
initial cameras, I compute the stationary distribution of camera formats based on the observables
in the first period, as in Hendel and Nevo (2006).31
5.10.6
Dimensionality reduction
Dimensionality reduction follows Equation (10), where I impose that the index function of year
dummies – which summarizes the role of yearly technology on the evolution of picture quality – follows a linear transformation at the point of camera switching. This in spirit follows
Gowrisankaran and Rysman (2012), who assume that the value function is a Markov index function.
Different from their work, however, I can compute the index Yit given parameters φy – as it is an
explicit function of the primitives – and hence avoid the extra fixed point loop in Gowrisankaran
and Rysman (2012). Therefore, the algorithm in estimation does not deviate from the classical
Nested Fixed Point algorithm in Rust (1987).
5.10.7
Discount factor
Finally, I give all consumers a discount factor of 0.95 monthly. The discount factor implies that the
consumers will discount away 70% of the value of a camera in two years, which I find intuitive.
This also implies an annual discount factor of 0.54, which is lower than the field-data estimates by
31 Alternatively,
one could model the initial brand-format distributions. I only model the initial camera format
distributions because, monthly choice probability being close to zero, the brand-format choice probability matrix is
more likely to be singular at some parameter values.
31
Table 4: Production function estimates
parameter
Return to human capital - all compacts
2.30
- all DSLRs
2.62
Year of intro - 2008
-0.29
- 2009
-0.31
- 2010
-0.23
- 2011 or later
0.12
Scale of error term (σν )
0.68
s.e.
0.26
0.20
0.01
0.04
0.07
0.07
0.01
Note: This table reports structural estimates for the production function. Year dummies before 2008 is restricted to
zero because of potential collinearaity with the individual-specific constant. Bootstrap standard errors are reported,
which are computed from estimates of 20 random samples with replacement.
Dubé et al. (2009) (0.7), but higher than the estimates for the non-durable goods case in Yao et al.
(2012).32
6
Estimation results
6.1
Production function
Table 4 shows that the returns to experience is 13.9% higher, when using a DSLR camera – which
qualitatively confirms the pattern shown in Appendix Figure 12. This shows that an advanced camera is a strong complement to consumer human capital (and vice versa). However, quantitatively,
the model finds that a DSLR camera does not produce as high the picture quality as the descriptive
evidence, because of the endogenous camera choices made by individuals of different experience.
6.2
Initial picture quality
I estimate the constant term in the production function, qi , separately for each type of consumers.
I interpret them to be the initial picture quality prior to registering on Flickr. Among other parameters, Table 5 presents estimates of qi , which is −2.88 for the first type, and −0.96 for the second.
This indicates a small part of the sample (34%) registered their Flickr accounts (i.e., enters the
32 The discount factor inYao et al. (2012) is close to zero annually, but is reasonable in their context of mobile phone
contracts, since it requires a much shorter-term thinking on the consumer side.
32
Table 5: Utility, learning and switching cost estimates
Utility: pref to quality (αi )
- effort (ei )
- expenditure (βi1 )
- expenditure squared (βi2 )
- preference to DSLR (λDSLR )
- preference to Canon (λCanon )
- preference to Nikon (λNikon )
- switching formats (λFormatSwitch )
- switching brands (λBrandSwitch )
Initial picture quality (qi )
Shape parameter (κi )
)
Switching cost: baseline (sbaseline
i
f ormat
- additional from across formats (si
)
brand
- additional from across brands (si
)
Type probability
"novices"
1.92
0.01
-1.34
0.06
3.81
-1.35
-1.31
-0.50
-1.82
-2.88
0.08
0.06
0.07
0.07
0.66
s.e. "experts" s.e.
0.26
1.28 0.27
0.11
-0.88 0.61
0.31
-1.34 0.29
0.02
0.08 0.02
1.42
3.19 1.19
0.20
-1.57 0.28
0.23
-1.67 0.17
0.19
-0.75 0.22
0.22
-2.46 0.15
0.24
-0.96 0.34
0.02
0.05 0.02
0.03
0.08 0.04
0.02
0.10 0.03
0.03
0.11 0.02
0.09
0.34 0.09
Note: This table reports structural estimates for the rest of the parameters. Bootstrap standard errors are reported,
which are computed from estimates of 20 random samples with replacement.
sample) with considerably more photography experience compared to the rest. For this reason, I
denote this smaller segment the “experts”, and the rest the “novices”. Also, note that the acrossconsumer difference is much larger than the average within-consumer evolution of picture quality,
the latter shown in Figure 6.
6.3
Learning by doing
Table 5 also presents estimates for the returns to learning and the switching cost. The shape parameter κi determines the importance of consumer human capital in shaping her picture quality,
i.e. the returns to learning. Also, the same parameter captures her learning speed, or the marginal
increase in knowledge from each time she takes pictures. From the estimates, I find that κi is far
below 1, which is strong evidence for decreasing returns to experience. In addition, the returns to
learning implied by the shape parameter is different across segments, and I find that for the novice
consumers, additional experience gained in the sample plays a much larger role, than that of the
experts. This indicates that those with lower initial knowledge stocks learn faster – consistent with
the decreasing marginal returns from learning.
33
Given the parameter estimates, Figure 8 presents the evolution of consumers’ (maximum) picture quality in monetary value, conditional on the experience in photography.33 Given the consumer type and the camera format, the figure clearly shows learning by doing with decreasing
speed. Quantitatively, the first year of experience increases the value of a novice’s picture quality
by a monetary value of $60. This is to say, for a novice consumer, she now values the picture-taking
activity in a month by 60 US dollars more. By the end of year 5, this value is further increased to
$120. On the other hand, the contribution of experience for an expert consumer is much lower, due
to the lower returns to learning (κ). The first year of experience increases picture quality by $30,
while the first 5 years increases it by $50.34
The returns to experience found here seem small at the first glance – especially compared to
the price differences between entry level and advanced digital cameras. However, note that the
accumulation of human capital is persistent, and hence, the utility gains from learning hence go far
into the future.
Comparable estimates include Shaw and Lazear (2008), who find that the output of production
workers increases by 53% in their first 8 months on the job.
6.4
Non-transferable human capital and the switching cost
As mentioned in Section 5.10, I allow switching costs to be specific to within/across formats and
brands, imposing that the across- format or brand switching cost is no smaller than the withinformat or brand counterpart.
It is natural to expect that previous knowledge is less applicable if a user switches to a camera
of a different format, e.g. from a compact camera to a DSLR – as the handling and operation of
the product is very different, despite general principles of photography still apply. Moreover, I also
find that across-brand switching is an important source of switching cost. This implies that there
are vast differences in product design across different brands – such as brand-specific menu and
button layouts.
Table 5 also presents the switching cost estimates, separately for each segment. I find that for
33 Due to the vast differences in the initial conditions, I present the increases in the monetary value of picture
quality, relative to the first period.
34 In the model-free evidence section, the left panel of Figure 6 predicts faster increase in the picture quality, because
changes in the camera format are not controlled for.
34
novice with a DSLR
expert with a DSLR
150
150
value of picture ($; relative to first period)
value of picture ($; relative to first period)
picture quality in 2005 USD
quality after switching
100
50
0
0
2
4
years (0 = first picture)
100
50
0
6
0
2
4
years (0 = first picture)
6
Figure 8: Evolution in the monetary value of picture quality
Notes: The solid lines depict the evolution of the individuals’ monthly highest picture quality, in monetary values,
conditional on the years of experience. The dashed lines represent the monetary values of the instantaneous drop of
picture quality, at an across-brand, within-format camera switch. These are plotted separately for the novice and the
expert segments, only for consumers who use a DSLR. One can calculate the values for compact camera users by
multiplying γcompact /γdslr = 0.88.
35
a current “novice” consumer using a Canon compact camera, switching to another Canon compact
camera costs 6% of her human capital, while switching to a Nikon compact camera costs 13%.
If she decides to upgrade to a DSLR camera, switching to a Canon DSLR costs 13% of her human capital, and 19% for switching to a Nikon DSLR. For an “expert” consumer, the additional
knowledge she gains in sample is much more product specific; and for her, switching from a Canon
compact camera to a Nikon DSLR costs 26%.
The dashed curves in Figure 8 show the picture quality after a counterfactual, within-format
and across-brand switch (e.g. from a Canon DSLR to a Nikon one), in terms of monetary value
– which demonstrates the extent of the switching cost. Although the model for switching cost
is restrictive, the predicted, proportional switching cost patterns closely resembles the model-free
patterns in Figure 6, or in Appendix Figures 13.
6.5
Other utility parameters
Between segments, the utility parameters from consuming picture quality and from purchasing
cameras are different. A consumer from the novice segment derives slightly higher utility from
consuming a unit of picture quality, yet much smaller dis-utility taking pictures. The dis-utility
term, called “effort”, rationalizes that a consumer does not take pictures every period. It also
represents how costly it is for each type of consumers to obtain an additional unit of human capital.
Despite the higher “effort” for the expert segment, their low returns to learning and high initial
quality results in a flatter picture-taking decision and camera-demand decision profiles, compared
to the novice segment.
The nonlinear price effects show that, for a one-dollar price change, the individuals are much
more sensitive at the lower price range. For the two segments, they become insensitive to price
changes at 1,100 and 840 dollars, respectively. This covers all the observed compact camera prices,
and 98% of the average monthly DSLR prices. Hence, the model generates downward-sloping
demand.
The instantaneous utility parameters from camera purchase and brand switching – that are
unrelated to picture quality – show that there is considerably positive utility from purchasing a
DSLR camera. This might represent the utility from using the advanced camera features from
36
Table 6: Average short-run elasticities
(1) Canon Compact
(2) Nikon Compact
(3) Other Compact
(4) Canon DSLR
(5) Nikon DSLR
(6) Other DSLR
(1)
-1.8530
0.0062
0.0050
0.2032
0.0979
0.1517
(2)
0.0105
-1.8667
0.0051
0.2213
0.1012
0.1603
(3)
0.0102
0.0059
-1.8642
0.2070
0.0990
0.1574
(4)
0.0103
0.0058
0.0050
-3.5078
0.1001
0.1581
(5)
0.0101
0.0059
0.0051
0.2124
-3.7606
0.1657
(6) No purchase
0.0101
0.0101
0.0057
0.0056
0.0051
0.0049
0.2063
0.2047
0.1014
0.0962
-3.6410
0.1506
Note: This table reports short-run elasticities. I compute elasticities by first calculating the implied choice probabilities
for each type of consumer, and then the counterfactual choice probabilities when prices for a given brand-format in a
row are temporarily reduced by 10% for the given month. Then, elasticities are computed from the averaged choice
probabilities. For example, the first row, second column reads: a 10% temporary decrease in the price of Canon
compact cameras decreases the demand for Nikon compact camera by 0.105%.
these cameras, or simply from status effects of using the DSLRs, and rationalizes the tendency to
upgrade despite at a low human capital level. Finally, the utility parameters on brand-switching
and format-switching are conventionally negative, and captures alternative explanations to state
dependence that are unrelated to learning by doing.
6.6
Implied short-run price elasticities
To verify whether the model produces conventional price effects, I simulate price elasticities from
an instantaneous 10% price decrease for a given brand-format, which is not expected to last beyond
the given month. I calculate the implied choice probabilities for each type of consumers, with or
without the price change, and weight average the choice probabilities to compute the implied
demand. The price elasticities are then computed from the percentage changes in demand, as
responses to the 10% decrease in the prices.
Shown in Table 6, I find that the short-run price elasticities are conventional, as in other empirical demand estimation literature in the digital camera industry (Song and Chintagunta, 2003;
Gowrisankaran and Rysman, 2012). For example, a 10% decrease in the prices for Canon DSLRs
increases the product’s current-period demand by 35%. Most of the additional demand comes from
the consumers who would otherwise not purchase in this period (the “no purchase” category).35
35 On
average, the “no purchase” alternative has a baseline market share of 95%.
37
7
Counterfactual Experiments
7.1
Overview
With the parameter estimates, this section investigates 3 managerial implications. The first, and
perhaps most straightforward, implication is that consumer experience increases their tendency to
purchase advanced products. I find that for starting consumers, a 1-year increase in their human
capital raises the sales of the advanced cameras in the same brand by 26%. Secondly, since experience accumulation leads to higher utility from product usage, consumers directly demand the
provision of experience. I find that, for a 1-year increase in human capital, an average starting consumer is willing to pay $405 one-off, which is close to her expected discounted lifetime expenditure on digital cameras. Finally, since the imperfect transferability is most apparent in across-brand
switching, the switching cost in knowledge creates brand loyalty, making across-brand switching
increasing costly.
7.2
Learning by doing and sales of advanced products
Product usage becomes more enjoyable, when consumers are equipped with more experience on
the current products. An increase in the consumer human capital raises her valuation on the complementary product features – especially those in the advanced products – and downplays the
importance of price. In this experiment, I simulate the effect on sales, when consumers are given
1 extra year of experience, that can be fully applied to all brand-format of cameras. The results are
presented in Table 7, in the format of relative changes in demand.
I find that, for Canon compact camera users who are in the sample for a year, a 1-year human
capital shock increases their demand for Canon DSLR cameras by 26%, and Nikon DSLRs by
35%. The higher percentage increase for Nikon is due to the low base choice probabilities. At the
same time, it also increases the sales of compact cameras – by 21% and 28%, respectively. This
indicates that having a higher level of photography human capital complements the demand for
digital cameras, so that an individual will be more likely to upgrade her entry-level product to an
advanced one, or simply repurchase the same camera format to stay with the current technology.
For more experienced consumers, the effects are smaller or are negative. This is the result of
38
Table 7: Sales counterfactuals for Canon compact camera users
1 year from start
2 years
3 years
4 years
5 years
Canon compact Nikon compact Canon DSLR Nikon DSLR
1.2124
1.2816
1.2588
1.3549
1.0402
1.1257
1.1270
1.1378
0.9322
1.0443
1.0867
0.9487
0.8936
1.0302
1.0655
0.8912
0.9160
1.0277
1.0502
0.9022
Note: This table reports percentage changes in sales for consumers with different amount of experience, who are,
in the counterfactual experiment, given 12 months of additional human capital. For example, the first number reads:
for consumers who are in the sample for 1 year, increasing their human capital by 1 year increases their purchase
probability of Canon compact cameras by 21.24%, relative to the benchmark value.
the decreasing learning speed and the increasing switching cost.
7.3
The demand for experience
Having a larger human capital stock not only changes demand, but the direct impact on productusage utility also increases welfare (Michael, 1973). Hence, there is a direct demand for usage
experience that the supply side can provide. In the context of digital cameras, examples of such
include free product training,36 photo contests (to incentivize product usage), and so forth.
In the second counterfactual experiment, I measure the size of such demand, by the utilityequivalent monetary amount to a 1-year increase in experience. Specifically, I calculate the amount
of a fixed monthly subsidy, which provides the same expected sum of current and future utility as
a 1-year increase in the consumer’s human capital. Formally, for consumer i at time t, I find EVi
charged under any choice (including the outside option Bit = 0, which had zero expenditure), such
that
h
i
h
i
Eit max Uia (Sit ; Pk0t 1 (Bit 6= 0) − EVi , Hit ) − Eit max Uia (Sit ; Pk0t 1 (Bit 6= 0) , Hit + ∆H) = 0,
a
a
where Pk0t 1 (Bit 6= 0)−EVi characterizes expenditure after subsidy EVi while Hit +∆H is the human
capital after the increase in experience (∆H = 12 months). Finally, I present the discounted sum of
the fixed monthly payment, at δ = 0.95.
36 For example,
in India, Vietnam and potentially some other countries, Canon provides free short-lectures for users
who just purchased their DSLRs.
39
Table 8: Valuation for additional 1 year of experience
1 year from start
2 years
3 years
4 years
5 years
WTP: novice WTP: expert WTP: average monthly expenditure
447.4989
325.3024
405.9334
332.8261
354.9264
181.1099
295.8021
382.8253
236.3609
119.8353
196.7243
426.4285
181.7395
91.5023
151.0450
444.2696
161.2669
82.4507
134.4573
464.0309
Note: This table reports, for consumers in different segments and with different levels of experience, the amount of
compensation that is welfare-equivalent to a counterfactual increase in human capital by 12 months. I compute the
amount of a fixed monthly subsidy, that is equivalent of this policy change – in the sense that it equates the expected
sum of future utility for the consumer. I then take discounted sum of this stream of subsidy. For example, the first
number reads: for consumers with 1 year of experience, her valuation of the 1-year extra human capital is measured as
the utility-equivalent one-off tax at 447 dollars. The last column reports the expected discounted lifetime expenditure
for consumers with the corresponding experience (without the extra human capital). This number is provided as a
benchmark to understand the magnitude of consumer valuation of knowledge.
Table 8 presents the results, separately for consumers from different segments, with different
experience levels. I find that across segments, the valuations for additional experience are comparable. This is because although the expert segment has lower returns to experience in the production
function, it is also more costly for them to obtain experience on their own, because of the higher
ei . On the other hand, the less experienced consumers value the additional human capital much
higher than the experienced consumers, due to the decreasing marginal returns to learning. An
average starting consumer values the additional experience by $405, which is even higher than
their discounted expected lifetime expenditure in the digital camera product category. As a comparison, the tuition fee for a one-month New York Institute of Photography online course is $50,
yielding a discounted sum of $460 for a year.37 For the more experienced consumers, the valuation decreases due to decreasing returns to experience – down to $134, or 29% of the consumer’s
expected lifetime expenditure.
7.4
Product specific experience and brand loyalty
The final two experiments focus on the effect of non-transferable consumer human capital, where
the across-brand component plays a major role. For example, switching from a Canon compact
37 Information
from http://www.nyip.edu/courses/professional-photography in September 2014.
40
camera to a Nikon DSLR costs 17% of a novice consumer’s human capital stock, or 26% of an
expert’s. Switching within the same brand would have costs the two consumers 13% or 17%
of their human capital stock, respectively. This difference between within- and across- brand
switching cost incentivizes the consumer to stick to a particular brand, and the effect is accentuated
when experience with the brand further accumulates. As a result, not only would a consumer avoid
brand switching ex post, in fear of losing part of the already accumulated experience, but they also
tend to avoid brand switching ex ante, by planning their brand choice ahead in accordance with
the expected long-run product characteristics. These two aspects are shown in two counterfactual
experiments.
To demonstrate that experience accumulation can cause the consumer to be locked in, I simulate
counterfactual choice probabilities when the brand-switching cost component in human capital is
removed. Table 9 presents, for consumers who are currently using Canon compact cameras, the
relative changes in choice probabilities for both entry-level and advanced products the under this
counterfactual scenario. I condition on that the consumers are currently using Canon compact
cameras, to better see the differences across formats and brands.
I find that for consumers with more experience, switching costs are more and more salient in
changing their choice profiles. For consumers with 5 years of experience, the sales of Nikon DSLR
cameras and Canon compact cameras increases by 28% and 25%, respectively. This naturally
comes from the increasing nature of switching cost in the consumer knowledge stock. From the
table, those who will switch to DSLRs at some point are more willing to switch to Nikon DSLRs,
while those who decides to stick to the compact cameras (due to preference as well initial human
capital heterogeneity), decides to increase their shares of expenditure on Canon. The second point
seems counter-intuitive, but comes from the high across-brand switching dis-utility which is still
present. Hence, with higher potential human capital (due to the switching cost decrease), some
consumers will increase their purchase probability into compact cameras, but will do so within
brand.
Appendix Table 3 shows the results of another experiment, where I keep the switching costs
as estimated, but eliminate the brand-switching dis-utility, in the fifth row of Table 5. This can
be interpreted as “other switching costs” that are unrelated to consumption of picture quality. I
find that elimination of brand-switching dis-utility increases the purchase probabilities of Nikon
41
Table 9: Choice probability changes without switching cost, Canon compact camera users
1 year from start
2 years
3 years
4 years
5 years
Canon compact Nikon compact Canon DSLR Nikon DSLR
0.0027
0.0040
0.0023
0.0039
0.0088
0.0111
0.0060
0.0148
0.0647
0.0069
0.0180
0.0599
0.1484
0.0119
0.0284
0.1613
0.2542
0.0312
0.0363
0.2813
Note: This table reports counterfactual changes in the predicted choice probabilities when all switching costs in human
capital transition are eliminated. The changes are reported as relative to the benchmark model estimates, averaged
across all consumers in the sample, conditional on their level of experience. For example, the first number reads: the
demand for Canon compact cameras will increase by 0.27%, when all switching costs in human capital transition are
eliminated.
DSLR cameras by up to 9 times. Even though this seems to be a much larger effect, it is not clear
whether the parameters we estimated are truly causality (versus merely heterogeneity in the brand
preferences).
On the other hand, expectations of consumer switching costs also lead to pre-planning on
brand choice, in the sense that because choices of brand can become increasingly costly over time,
consumers put great importance on the long-run expected product characteristics, when making
their brand choices. In my final counterfactual experiment, I show that consumers instantaneously
switch to the brand with more attractive permanent prices. In fact, the effect is large even for
consumers who are not very forward-looking, as in the structural analysis, they have a discount
rate that is drastically higher than the market interest rate.38
In Table 10, I simulate price elasticities from a known, permanent 10% price decrease – the
only difference from Table 6 is that the price changes are permanent and the consumers rationally
expect this.39 Most notably, the table shows that, in face of a permanent price change in one of the
DSLRs, many consumers will instantaneously switch to the compact cameras in the same brand.
This is because within-brand switching costs are much lower than across-brand switching costs,
hence making knowledge acquired from using the previous product much more valuable if the
consumer sticks with the same brand. As a result, when Nikon DSLRs are permanently cheaper, a
Canon user who is not prepared to purchase a DSLR right away, might decide to purchase a Nikon
38 I
assume a discount factor of 0.95 monthly, or equivalently, 0.54 annually.
only report elasticities for Canon and Nikon products for the ease of reading, while the complete table can be
found at Appendix Table 2.
39 I
42
Table 10: Average long-run elasticities for Canon and Nikon products
(1)
(2)
(4)
(5)
(1) Canon Compact -1.9383 0.0288 0.0268 0.0417
(2) Nikon Compact 0.0126 -1.9628 0.0245 0.0093
(4) Canon DSLR
-1.6431 0.8915 -3.9841 1.1946
(5) Nikon DSLR
0.3735 -2.3449 0.7000 -4.8753
Note: This table reports long-run elasticities, computed from demand responses to permanent price changes of 10%
for a given brand-format of camera. For example, the first row, second column reads: a foreseeable, permanent 10%
price decrease for Canon compact cameras decreases the demand for Nikon compact cameras, by 0.427%. Only
Canon and Nikon products are included for the ease of reading, while the complete table can be found at Appendix
Table 2.
compact camera. In doing so, her further-acquired knowledge can be transferred within Nikon,
with lower attrition. Analogous arguments hold for Nikon users in this situation. This explains
that, intertemporally, products within a brand can be strong intertemporal complements.
8
Concluding Remarks
This paper quantifies the importance of consumer learning by doing – i.e. accumulation of productspecific human capital through usage – on their demand for advanced products. In the context
with entry-level and advanced digital cameras, I measure the returns to consumer experience, via
looking at how a homogenous set of viewers receive a consumer’s pictures, taken at different points
in time. On the one hand, experience leads to higher utility from product usage (in this case, via
higher picture quality). Thus, not only is the provision of knowledge valued by the consumer, but it
also increases her demand for advanced products. I find that for a beginner, increasing her human
capital by 1 year results in a 26% higher demand for within-brand product upgrades. On the other
hand, I find that up to 26% of a consumer’s product experience is not fully transferable, and this
discourages product switching – in particular between brands where product-design differences
are greater. As a result, more experienced consumers display greater brand loyalty; and knowing
so, even mildly forward-looking consumers consider products across brands to be much more
substitutable in the long run, than those within a brand.
The model of consumer learning by doing that is proposed in this paper has great generalizability in home electronics, sports equipment, entertainment, and other categories that require
43
consumer skills to use the products. From a managerial point of view, understanding the evolution
of consumer knowledge not only helps understand the evolution of their demand – in particular the
migration from entry level to more advanced products, but it also helps understand their tendency to
be locked in to products that are similarly designed as their previous ones. Further, because usage
experience is desirable on its own, there is demand for supply-side provision of consumer knowledge, such as the firm offering training services, competitions in user content creation, or simply
designing products that are easy to use. From the manager’s perspective, whether such actions are
profitable depends not only on the returns to experience, but also on how widely-applicable the
product knowledge is.
The empirical exercise in this paper is done on a relatively small sample, which might be
non-representative; however, the difficulty of measuring the returns to experience limits the usage
of more standard market share data, used in Song and Chintagunta (2003), Gowrisankaran and
Rysman (2012), among others. This is related to the general difficulty of identifying the source of
state dependence from choice data alone (Ching et al., 2013), which is beyond the scope of this
paper.
44
References
Adner, Ron and Daniel Levinthal (2001), ‘Demand heterogeneity and technology evolution: implications for product and process innovation’, Management science 47(5), 611–628.
Alba, Joseph W and J Wesley Hutchinson (1987), ‘Dimensions of consumer expertise’, Journal of
consumer research pp. 411–454.
Albuquerque, P. and Y. Nevskaya (2012), ‘The impact of innovation on product usage: A dynamic
model with progression in content consumption’.
Becker, G.S. (1965), ‘A theory of the allocation of time’, The economic journal 75(299), 493–517.
Besanko, David, Ulrich Doraszelski and Yaroslav Kryukov (2014), ‘The economics of predation:
What drives pricing when there is learning-by-doing?’, American Economic Review 104, 868–
897.
Besanko, David, Ulrich Doraszelski, Yaroslav Kryukov and Mark Satterthwaite (2010), ‘Learningby-doing, organizational forgetting, and industry dynamics’, Econometrica 78(2), 453–508.
Bronnenberg, Bart J, Jean-Pierre H Dube and Matthew Gentzkow (2012), ‘The evolution of brand
preferences: Evidence from consumer migration’, American Economic Review 102(6), 2472–
2508.
Ching, Andrew T, Tülin Erdem and Michael P Keane (2013), ‘Invited paper-learning models: An
assessment of progress, challenges, and new developments’, Marketing Science 32(6), 913–938.
Clarkson, Joshua J, Chris Janiszewski and Melissa D Cinelli (2013), ‘The desire for consumption
knowledge’, Journal of Consumer Research 39(6), 1313–1329.
Crawford, G.S. and M. Shum (2005), ‘Uncertainty and learning in pharmaceutical demand’,
Econometrica 73(4), 1137–1173.
Dubé, Jean-Pierre, G Hitsch and Pranav Jindal (2009), Estimating durable goods adoption decisions from stated preference data, Technical report.
45
Dubé, J.P., G.J. Hitsch and P.E. Rossi (2010), ‘State dependence and alternative explanations for
consumer inertia’, The RAND Journal of Economics 41(3), 417–445.
Erdem, Tulin and Michael P. Keane (1996), ‘Decision-making under uncertainty: Capturing
dynamic brand choice processes in turbulent consumer goods markets’, Marketing Science
15(1), 1–20.
Erdem, Tülin, Michael P Keane, T Sabri Öncü and Judi Strebel (2005), ‘Learning about computers: An analysis of information search and technology choice’, Quantitative Marketing and
Economics 3(3), 207–247.
Foster, A.D. and M.R. Rosenzweig (1995), ‘Learning by doing and learning from others: Human
capital and technical change in agriculture’, Journal of Political Economy pp. 1176–1209.
Gowrisankaran, Gautam and Marc Rysman (2012), ‘Dynamics of consumer demand for new
durable goods’, Journal of Political Economy 120(6), 1173–1219.
Guadagni, Peter M and John DC Little (1983), ‘A logit model of brand choice calibrated on scanner
data’, Marketing science 2(3), 203–238.
Hendel, Igal and Aviv Nevo (2006), ‘Measuring the implications of sales and consumer inventory
behavior’, Econometrica 74(6), 1637–1673.
Jovanovic, B. and Y. Nyarko (1996), ‘Learning by doing and the choice of technology’, Econometrica: Journal of the Econometric Society pp. 1299–1310.
Kasahara, H. and K. Shimotsu (2009), ‘Nonparametric identification of finite mixture models of
dynamic discrete choices’, Econometrica 77(1), 135–175.
Keane, M.P. and K.I. Wolpin (1994), ‘The solution and estimation of discrete choice dynamic
programming models by simulation and interpolation: Monte carlo evidence’, The Review of
Economics and Statistics pp. 648–672.
Levitt, Steven D, John A List and Chad Syverson (2013), ‘Toward an understanding of learning by doing: Evidence from an automobile assembly plant.’, Journal of Political Economy
121(4), 643–681.
46
Magnac, Thierry and David Thesmar (2002), ‘Identifying dynamic discrete decision processes’,
Econometrica 70(2), 801–816.
Melnikov, Oleg (2000), ‘Demand for differentiated durable products: The case of the us computer
printer market’, Manuscript. Department of Economics, Yale University .
Michael, R.T. (1973), ‘Education in nonmarket production’, The Journal of Political Economy
pp. 306–327.
Murray, K. and G. Häubl (2005), ‘Explaining cognitive lock-in: The role of skill-based habits of
use in consumer choice’, Journal of Consumer Research, Vol. 34, No. 1, 2007 .
Nam, Myungwoo, Jing Wang and Angela Y Lee (2012), ‘The difference between differences:
How expertise affects diagnosticity of attribute alignability’, Journal of Consumer Research
39(4), 736–750.
Nelson, Phillip (1970), ‘Information and consumer behavior’, The Journal of Political Economy
78(2), 311–329.
Nosal, Kathleen (2012), ‘Estimating switching costs for medicare advantage plans’, Unpublished
manuscript, University of Mannheim .
Ratchford, Brian T (2001), ‘The economics of consumer knowledge’, Journal of Consumer Research 27(4), 397–411.
Rust, J. (1987), ‘Optimal replacement of gmc bus engines: An empirical model of harold zurcher’,
Econometrica: Journal of the Econometric Society pp. 999–1033.
Shaw, Kathryn and Edward Lazear (2008), ‘Tenure and output’, Labour Economics 15(4), 704–
723.
Song, Inseong and Pradeep K Chintagunta (2003), ‘A micromodel of new product adoption with
heterogeneous and forward-looking consumers: Application to the digital camera category’,
Quantitative Marketing and Economics 1(4), 371–407.
Wernerfelt, Birger (1988), ‘Umbrella branding as a signal of new product quality: An example of
signaling by posting a bond’, Rand Journal of Economics 19(3), 458–466.
47
Yao, Song, Carl F Mela, Jeongwen Chiang and Yuxin Chen (2012), ‘Determining consumers’
discount rates with field studies’, Journal of Marketing Research 49(6), 822–841.
Youn, N., I. Song and D. MacLachlan (2008), ‘A multi-category approach to modeling consumer
preference evolution: The case of sporting goods’.
48
Appendices
A
Returns to experience in photography
A.1
Overview
This section estimates the returns to experience in photography in reduced form, where the model
is flexible enough to allow for deviations from the assumptions used to infer picture quality, in
Section 3.3. Hence, this section also serves as a robustness check for the picture quality measure.
A.2
Specifications
I first further assume quadratic specification of picture quality qip from Equation (2):
2
qip = θ1 xip + θ2 xip
+ ∑ camit + qi0 + ηit
which is a quadratic on experience xip , plus a set of camera fixed effects, individual fixed effect
qi0 , and a error term ηit . I then regress
log (viewsip ) =
∑ Φt0t1 + zipψ + θ1xip + θ2xip2 + ∑ camit + qi0 + ηit ,
(13)
t0 ,t1
and the parameters θ1 and θ2 capture the returns to experience. Note that this specification shares
essentially the assumption I use to infer picture quality, other than the additional quadratic functional form on experience, and the separability in camera dummies.
There are, however, two potential concerns to the assumptions to Equation (2). First, the flow
of viewers could interact with experience, resulting in heterogeneous display-window effect. In
other words, Φt0t1 might be individual specific. The second concern is associated with the timing
of upload, i.e. the user might strategically choose the time to upload a picture based on its quality.
Both arguments point to the heterogeneity of the display window dummies.
With this in mind, I also estimate the returns to experience on a more-flexible specification.
Although this cannot be used to infer picture quality, it serves as an robustness check. Specifically,
49
Appendix Table 1: Returns to experience in photography
individual fixed effect individual-display window fixed effect
experience (100 months)
0.641***
0.546***
(0.012)
(0.015)
experience sq (0000s)
-0.139***
-0.116***
(0.005)
(0.007)
camera dummies
Yes
Yes
topic dummies
Yes
Yes
upload order
Yes
Yes
display window
Yes
No
months since joined Flickr
Yes
No
number of pics uploaded
Yes
No
Rsq.
0.112
0.006
obs.
1557232
1560224
Note: This table provides reduced form estimates of the returns to experience in photography. The dependent variable
is log of cumulative number of views, per picture level. The first column corresponds to Equation (13), where we infer
the returns to experience using within-individual variation, but adding covariates to control for aggregate calendar
time trend in Flickr. The second column reports estimates using within-individual-upload-time variations, based on
the specification in Equation (14). Measured in picture quality, the first specification estimates a 3-year return to
experience of 21.3%, or an annualized 6.7%; the second specification estimates a 3-year return of 18.2% – annualized
to 5.8%. Within-effect R-squared are provided.
I allow for interactions of individual heterogeneity and the display-window effects, resulting in
individual-display-time dummies Φ̃i,t0 ,t1 . Equation (13) now becomes:
2
log (viewsip ) = z̃ip ψ + θ1 xip + θ2 xip
+ ∑ camit + Φ̃i,t0 ,t1 + η̃it ,
(14)
where Φ̃i,t0 ,t1 now captures a combined effect of baseline picture quality qi0 and individual-specific
flow of viewers. We can regress (13) controlling for individual-batch fixed effects.
A.3
Estimates
The first column in Table (1) presents the estimation results from Equation (13). I find that the
returns to experience is positive within sample period, with a decreasing marginal return. This
is consistent with the learning curve measures in Shaw and Lazear (2008), Besanko et al. (2010),
Levitt et al. (2013),
among others. Quantitatively, the annual return to experience in the first 3 years
amounts to an increase in picture quality, such that it generates 6.7% more views.
50
I further check the sensitivity of this result to potential endogeneity problems, as discussed.
Shown in Column 2, the robust learning speed estimates are not economically different from the
benchmark estimates – hence, the inferred learning curve economically robust to the potential
concern of endogeneity.
B
A simple illustrative model
B.1
Model
This section provides a simple illustrative model to highlight the core mechanism. Assume that the
individual derives utility from the camera quality γk , k = h, l with γh > γl , and the human capital
Hit , which evolves exogenously as in Hit+1 = Hit + 1. Further, camera quality and human capital
enter multiplicatively into the utility, hence they are complements.
The individual starts at technology l, and makes a one-shot upgrade decision, Dit = 0, 1, at the
financial cost of β · P, as well as a proportional human capital cost s. Once upgraded, she cannot
turn back (and it is not rational of her to do so).
Finally, there is a random utility shock εidt , d = 0, 1, conditional on the decision. We can
summarize the assumptions by the utility function
uit = γk Hit − β P · 1 (Dit = 1) + εidt .
B.2
Static case
If the individual is myopic, her choice probability is
Pr (Dit = 1) = F (γh (1 − s) Hit − γl Hit − β · P)
= F ((γh − γl ) Hit − sγh Hit − β · P)
where F is the CDF of ∆ε. Clearly, if
(γh − γl ) Hit − sγh Hit > 0
51
or
γh /γl > (1 − s)−1
then the upgrade probability profile should be increasing in H. In words, when the consumer is
myopic, if the instantaneous gain in technology outweighs a function of the loss in skills, then the
individual tends to upgrade late due to future human capital growth.
B.3
Dynamic case: numerical results
When the individual maximizes the discounted sum of future utility flow, the analytical problem
becomes more complicated, due to the value function being hardly tractable. Instead, I provide
numerical results, which demonstrates that qualitatively the same property holds as in the static
case.
I set γh = 1.1, γl = 1, P = 12, and compare the resulting choice probability profiles when the
consumer is myopic (δ = 0), or δ = 0.95, given that there is no switching cost (s = 0), or s = 0.3.
To make choice probabilities comparable across static/dynamic cases, I set β = 0.2 in the static
case, and β = 3 in the dynamic case.
Figure 9 shows the results. The two figures in Panel A confirms the previous analysis: when
switching cost is large, upgrade probabilities will be decreasing in human capital. However, compared with Panel B, which shows the results of a dynamic model, there are two conclusions to be
drawn:
(1) Choice probabilities still can be decreasing in human capital at high switching costs – and
the qualitative take-away from the illustrative model holds in the dynamic case;
(2) However, when forward looking, the consumers are able to “shift” the upgrade timing,
and upgrade earlier due to the lower switching cost. In the dynamic case (Panel B, compared to
Panel A), the area below the choice probability profiles – i.e. the fraction of population using an
advanced camera – is much less sensitive to the increase in switching cost.
52
B.1: δ=0.95, β=3, s=0
1
0.8
0.8
upgrade probability
upgrade probability
A.1: δ=0, β=0.2, s=0
1
0.6
0.4
0.2
0
0
5
10
0.6
0.4
0.2
0
15
0
5
experience
0.8
0.8
0.6
0.4
0.2
0
5
15
B.2: δ=0.95, β=3, s=0.3
1
upgrade probability
upgrade probability
A.2: δ=0, β=0.2, s=0.3
1
0
10
experience
10
0.6
0.4
0.2
0
15
experience
0
5
10
15
experience
Figure 9: Numerical choice probability in various cases
Notes: The four figures are numerical choice probabilities along consumer experience, given different parameter
settings. Panel A depicts choice probabilities under a static model, with or without switching cost; while panel B
depicts that of a dynamic model.
53
C
An alternative model of household production
One can derive an alternative structural model where an advanced product is more complicated
than an entry-level product, and hence, using it requires more effort. Here, the role of consumer
human capital reduces such effort cost, and allows one to spend more effort into picture taking. On
the other hand, switching to a different product makes some of the human capital obsolete, which
increases the effort cost required to use the new product.
This section shows that such a model generates similar predictions to our benchmark model.
First, assume that expected picture quality is produced by the camera technology specific to camera
k, Γk , and some consumer effort, Eit :
E [Qikt ] = Γk · Eit
but spending effort incurs a cost, specific to the current camera, and dependent on the current
human capital:
ck (Hit ) · (Eit )σ
with ck (Hit ) > 0 and σ > 1. This is to say, it is always possible to produce the best picture, but for
a consumer with low human capital level, producing high quality picture might not be rational. In
addition, the marginal cost of producing additional picture quality is increasing.
The consumer maximizes the net utility from consuming the picture quality net of the cost. In
other words, she solves
max αΓk · Eit − ck (Hit ) · (Eit )σ
Eit
where again, α is the marginal utility from consuming the picture quality.
αΓk − ck (Hit ) · σ (Eit∗ )σ −1 = 0,
or
Eit∗
1
−1 σ −1
−1
= σ αΓk (ck (Hit ))
54
Substitute Eit∗ into the production function, and assume that
ck (Hit ) = c0k /Hit ,
and we can see the similarity to the benchmark model:
1
σ −1
E [Qikt ] |Eit∗ = Γk · σ −1 αΓk (ck (Hit ))−1
1
1
σ
1
1
σ
−
1
1
1
σ
−
1
1
= σ − σ −1 α σ −1 Γkσ −1 ck (Hit )− σ −1
1
= σ − σ −1 α σ −1 Γkσ −1 c0kσ −1 Hitσ −1 .
If one imposes that κ =
1
σ −1
and
γk = σ − σ −1 α σ −1 Γkσ −1 c0kσ −1
κ
α
1 + κ 1+κ
=
Γk ·
κ
c0k
then we have a similar specification to the benchmark model.
However, note that the interpretation is not completely the same. Here, κ retains a similar
interpretation – as a parameter that governs the returns to human capital. However, γk becomes
an index of the returns to camera format, and a ratio that captures the tradeoff between spending
effort and consuming high-quality pictures.
D
Model fit
This section examines the in-sample model fit of the purchase probabilities as well as picture
quality, in comparison with the data, separately by each segment. To do this, I first compute the
posterior type probability, and segment the raw data when the posterior probability of an individual
being in one type is greater than 0.9. Then, I plot the time trend in raw data and purchase probability
by type, together with the model predicted average quality and purchase probability.
Figure 10 shows that the model fits both purchase and picture quality data well – in that the
model fit almost always lies in the 95% confidence interval of the sample average. For the novice
55
type, the model slightly over-predicts picture quality around year 4-6; and for the experts, the
model under-predicts quality around year 2-3. As for choice probabilities, the model over-predicts
the purchase probability of a DSLR around year 6-8.
novice
expert
0.8
2.6
picture quality
picture quality
0.6
0.4
0.2
2.4
2.2
2
0
1.8
0
2
4
6
8
10
monthly purchase prob. of a DSLR
monthly purchase prob. of a DSLR
−0.2
data
model prediction
95% CI of data
0.1
0.05
0
0
2
4
6
8
years (0 = first picture)
10
0
2
0
2
4
6
8
10
4
6
8
years (0 = first picture)
10
0.1
0.05
0
Figure 10: Model fit: purchase probabilities and picture quality
Notes: These figures present in-sample fit for the baseline structural model, for picture quality and purchase probability
of a DSLR, separately for both types.
56
Additional Figures and Tables
DSLR to compact
9
9.5
10
number of pictures taken
1.2
8
8.5
max quality in a month
.9
1
1.1
9
10
11
number of pictures taken
max quality in a month
.9
1
1.1
1.2
12
1.3
compact to compact
10.5
E
0
1
years (0 = camera switching)
−1
0
1
years (0 = camera switching)
11
1.7
8
9
10
number of pictures taken
max quality in a month
1.3
1.4
1.5
1.6
11
8
9
10
number of pictures taken
max quality in a month
1
1.2
1.4
1.2
max picture quality
number of pictures taken
7
.8
max picture quality
number of pictures taken
−1
2
DSLR to DSLR
1.6
compact to DSLR
0
1
years (0 = camera switching)
8
.8
2
2
−1
0
1
years (0 = camera switching)
7
−1
max picture quality
number of pictures taken
7
.8
max picture quality
number of pictures taken
2
Figure 11: Switching cost controling for camera formats
Notes: These figures present monthly maximum picture quality for each individual, before and after camera switching,
conditional on the camera formats before and after. For detailed notes, see Figure 5.
57
current DSLR users
max picture quality in a month
1
1.2
.6
.6
.8
max picture quality in a month
.8
1
1.2
1.4
1.4
current compact camera users
0
1
2
3
4
years (0 = first picture)
5
0
1
2
3
4
years (0 = first picture)
5
Figure 12: Quality of best picture, conditional on camera format
Notes: These two figures present monthly maximum picture quality for each individual, against years of experience,
conditional on usage of a given format of camera. For detailed notes, see Figure 5.
current DSLR users
.8
.8
1
switching cost
1.2
switching cost
1
1.2
1.4
1.4
current compact camera users
0
1
2
3
4
years (0 = first picture)
> 3 months
5
0
<= 3 months
1
2
3
4
years (0 = first picture)
> 3 months
5
<= 3 months
Figure 13: Quality of best picture from new and familiar cameras
Notes: These two figures present monthly maximum picture quality for each individual, separately by the format of
cameras, and by whether the camera has been in use for more than 3 months. For detailed notes, see Figure 5.
58
Appendix Table 3: Choice probability changes without brand-switching dis-utility, Canon compact
camera users
Canon compact Nikon compact Canon DSLR Nikon DSLR
1 year from start
-0.1317
5.8351
0.4944
10.0064
2 years
-0.1599
5.2690
0.3815
8.6193
3 years
-0.1400
5.0750
0.3550
8.1983
4 years
-0.1108
4.9377
0.3605
8.3414
5 years
-0.0649
4.8322
0.3575
8.8907
Note: This table reports counterfactual changes in the predicted choice probabilities when the brand switching disutility term is set to zero. The changes are reported as relative to the benchmark model estimates, averaged across all
consumers in the sample, conditional on their level of experience. For example, the first number reads: the demand
for Canon compact cameras will decrease by 19.47%, when the brand-switching dis-utility is eliminated.
Appendix Table 2: Average long-run elasticities (complete table)
(1) Canon Compact
(2) Nikon Compact
(3) Other Compact
(4) Canon DSLR
(5) Nikon DSLR
(6) Other DSLR
(1)
-1.9383
0.0126
-0.0027
-1.6431
0.3735
-0.0728
(2)
0.0288
-1.9628
-0.0033
0.8915
-2.3449
-0.0825
(3)
-0.0059
-0.0232
-1.9165
0.2713
-0.2718
-1.0620
(4)
(5)
0.0268 0.0417
0.0245 0.0093
0.0186 0.0187
-3.9841 1.1946
0.7000 -4.8753
0.4714 0.4955
(6) No purchase
0.0404
0.0028
0.0226
0.0017
0.0167
0.0015
0.9689
0.0775
0.4822
0.0527
-3.5842
0.0633
Note: This table reports long-run elasticities, computed from demand responses to permanent price changes of 10%
for a given brand-format of camera. For example, the first row, second column reads: a foreseeable, permanent 10%
price decrease for Canon compact cameras decreases the demand for Nikon compact cameras, by 0.427%.
59